Summary
QWEDClient.verify_code() never sends code to the hardened server-side CodeVerifier when qwed_new is importable. Installing the qwed package makes it importable, and the GitHub Action image has it too. The SDK instead runs a local CodeGuard: an AST name blocklist of 8 function names and 7 modules. Code that the server classifies as unsafe (admission BLOCKED) comes back from the SDK as status="VERIFIED", is_verified=True, with the message "Code structure verified safe".
The server-side bypasses fixed in GHSA-6rxr-gq43-694g and GHSA-4phf-6h72-9xfg are therefore still reachable through the SDK and the Action's engine: code mode.
Affected components
| Component | Location |
|---|---|
QWEDClient.verify_code |
qwed_sdk/client.py L155-182: try: CodeGuard().verify_safety(...), with the API used only on ImportError |
CodeGuard |
src/qwed_new/guards/code_guard.py |
GitHub Action verify with engine: code |
action_entrypoint.py → client.verify_code(...) |
Root cause
CodeGuard.verify_safety flags only:
- direct
Namecalls or.attrcalls whose name is in{eval, exec, compile, open, system, popen, __import__, spawn}, and import/fromof{os, subprocess, sys, shutil, socket, pickle, pty}.
Everything else is reported as verified safe. That includes dynamic attribute lookup, builtins access through getattr, dunder/subclass traversal, importlib, string-built names and aliasing. The SDK treats this heuristic result as an authoritative VERIFIED.
Proof of concept (local, synthetic, inert payloads)
from qwed_new.guards.code_guard import CodeGuard
from qwed_new.core.code_verifier import CodeVerifier
from qwed_new.core.diagnostics import admission_decision
for code in ["getattr(__builtins__, 'ev'+'al')('1')",
"().__class__.__base__.__subclasses__()"]:
print(CodeGuard().verify_safety(code)["verified"], # True
admission_decision(CodeVerifier().verify_code(code)).value) # BLOCKED
Through the client, QWEDClient(api_key=..).verify_code(code) returns VerificationResult(status='VERIFIED', is_verified=True, ...) without making any network call.
Impact
Integrity of code admission. An application, agent framework or CI job that uses QWEDClient.verify_code (or the Action's code engine) to decide whether LLM-generated code may run will accept code the QWED server rejects. That code can reach eval/exec and arbitrary modules. Attacker capability: control over the code being verified (model output). No credentials are needed.
Affected versions
qwedon PyPI: 3.0.1 to 7.3.0. The local-firstCodeGuardpath was added infd863c8, first published in 3.0.1 (3.0.0 was not published to PyPI). Confirmed at runtime on 7.3.0.- GitHub Action
verifymode withengine: code: v7.1.0 to v7.3.0, where the Action gates onis_verified.
Patches
Fixed in 8.0.0 (88bda3d):
QWEDClient.verify_code(sync and async) always calls the server's/verify/code, andis_verifiedfollows the server's admission (GHSA-jppc-7r58-7pjh). A server that cannot be reached givesUNVERIFIABLE, neverVERIFIED.CodeGuardalso rejects reflective calls (getattr,setattr,globals,locals,vars),__builtins__, dunder attribute access, dotted imports, andimportlib,builtins,ctypesandmarshal. Its pass is markedadvisory; it is never reported as a verification.
Upgrade with pip install -U qwed (8.0.0 or later) and pin the Action to QWED-AI/qwed-verification@v8.0.0.
Workarounds
Call the server's /verify/code directly and gate on admission == "ADMIT", or use CodeVerifier().verify_code() with admission_decision() in process.
Related
GHSA-6rxr-gq43-694g, GHSA-4phf-6h72-9xfg (server CodeVerifier); GHSA-4v5r-g7f4-vvgc (the same pattern for SDK local math); GHSA-v8qq-v427-qrqc (QWEDLocal.verify_code). The SQL counterpart is GHSA-45jv-2j44-4c2q.
